Motion Vector based Abnormal Moving Vehicle Detection in Nighttime

Collision prediction during nighttime driving is helpful to avoid accidents because the driver has time to prepare upcoming situations. This research intends to find a real-time solution for frontal and lateral collision warning in an intelligent vehicle. Motion vectors of scene objects are estimated from time-varying frames. Small motion vectors are eliminated by using empirical threshold values pre-determined based on distribution of motion magnitudes. The remaining motion vectors are segmented to rectangular regions by using the unsupervised clustering K-means algorithm. After ROI setting, all segment candidates are classified into vehicle or non-vehicles by using SVM algorithm. From our experiments on real driving situations, the collision risk from lateral and preceding abnormal moving vehicles can be predicted with 91.96% accuracy. The detected abnormal moving vehicles include on-coming, lane change, abrupt speed change, roadside-parking, and overtaking.